Cloud Hosting

MHC-Bench

The MHC-Bench dataset consists of 2,232,937 peptides of length 9 and 115 MHC alleles. All the MHC alleles in this dataset are human MHC molecules (i.e. Human Leukocyte Antigens or HLA). Out of the 115 MHC alleles, about half are HLA-B, a third are HLA-A and remaining HLA-C. The MHC-Bench dataset contains 3,464,013 peptide-MHC pairs previously unseen by predictors during training. It is worth noting that the peptides by themselves may have been seen in the training data paired with a different HLA allele. The peptide overlap between training data for investigated predictors and MHC-Bench is shown in Table1. A description of the construction and composition of the dataset is presented in Methods section.

We evaluated the performance of four MHC class I predictors on the MHC-Bench dataset MHCflurry, MHCfovea, NetMHCpan and TransPHLA. The choice of the four predictors was guided by their popularity and performance reported in the literature1,2,4,47. MHCflurry2.047 is an ensemble predictor of 10 neural networks for predicting presentation of a peptide. It supports 14,993 MHC class I alleles. MHCflurry provides three scores, namely Binding Affinity (BA), Processing score and Presentation Score (PS). PS is produced by a logistic regression model combining the binding affinity and processing scores. Processing score captures the antigen probability which combined with binding affinity substantially improves the performance of the predictor47.

NetMHCpan4.11, an ensemble of 50 neural network models, produces Elution Ligand (EL) and Binding Affinity (BA) score and we refer to these modes as NetMHCpan-EL and NetMHCpan-BA respectively. It utilizes NNAlign_MA48, an artificial neural network that predict both BA and EL scores. For both modes, peptides with rank 2% or less are considered as binders and 0.5% or less as strong binders. It supports 11,000 MHC class I alleles.

MHCfovea2, is an ensemble of multiple CNN models that takes MHC allele sequence and peptide sequence as input to predict binding probability. In MHCfovea, ScoreCAM2,49 is applied to identify important positions and corresponding residues in input peptides and MHC allele sequence. It provides the motifs for the first and last 4 positions of the peptide for each allele along with the motif for MHC sequence.

TransPHLA4, is a transformer architecture that predicts binding probability for an input peptide and MHC allele sequence. Using the attention scores, important residues for each position of a 9mer can be obtain to generate a peptide motif for a given allele.

We use Area Under ROC (AUROC) and Area Under Precision-Recall Curve (AUPRC) as benchmark or performance metrics. Since peptide binding is MHC allele specific, we calculated the scores for each allele separately. The benchmark metrics for each allele are reported in Supplementary Data1, 2. The average benchmark metrics for the MHC class I predictors are reported in Table2. We find that all four predictors are comparable in their average performances across all alleles, as seen in Fig.2c. While the differences between the scores among predictors is minimal, we see that for most alleles, performance of MHCflurry-PS is marginally higher. Figure2a displays number of times (or number of alleles) the predictor was top performer based on AUROC and AUPRC score. We provide additional benchmarking analysis of the predictors in theSupplementary material (See Supplemental Fig.1 and Supplementary Note2) which indicates that the performance of the predictors are comparable across various metrics. The predictors achieve a higher performance on the AUROC metric, 0.950.98, as opposed to AUPRC where it is in the range of 0.750.86. For most alleles, there are fewer binding peptide-MHC pairs in comparison to non-binding peptide-MHC pairs. This is evident in the distribution plot in Fig.2b where for most of the alleles, only 110% of the peptides are binding. The %binding peptide-MHC pairs per allele is provided in Supplementary Data3). In imbalanced scenarios with few positive labels, the AUROC metric can be overly optimistic50. In contrast, AUPRC is relatively unaffected by this imbalance in the dataset owing to its focus on true positives (as true negatives are not into consideration). These two metrics were selected because they are used for benchmarking in the original paper describing the predictors. Recent work by Carrington et al.51 provides a useful discussion on use of AUROC and AUPRC metrics while also introducing a new metric called DeepROC.

a The number of alleles for which predictors exhibited the highest performances based on AUROC and AUPRC scores. b The percentage of binders in the MHC-Bench dataset per allele. Each dot represents a measurement for one allele. c Distribution of AUROC and AUPRC scores for the MHC class I predictors.

Explanations can be classified as either Global or Local. Global explanations for a predictor are distribution of input feature values across all inputs in the dataset. This offers a consolidated perspective on how the model utilizes input features for a specific output label. In contrast, local (or instance-based) explanations, focus on a single input - output instance. Local explanations typically show attribution of input features used for the prediction outcome. In the context of MHC class I predictors, binding motifs are examples of global explanations while a vector of attribution values for individual peptide positions forms a local explanation. It is worth noting that our work focuses on post-hoc explanations, i.e., explanations for existing models that have been previously trained. Post-hoc explanations are widely applicable as they can be used over models whose internal structure is not known or is too complex to interpret. Existing MHC class I predictors, like MHCfovea, focus on global explanations by generating binding motifs for MHC alleles. There is limited work on local instance-based explanations for this problem. In the next two sections, we motivate the need for local explanations for MHC class I predictors and discuss the additional information it can provide over global explanations.

An explanation for a 9mer peptide in our instance-based approach is represented as a length-9 vector of attribution values, generated through LIME or SHAP. Each positions attribution value can be positive or negative, with a positive (or negative) value indicating that the residue at that position contributes positively (or negatively) to the prediction. Our MHCXAI framework facilitates the generation of explanations for any MHC class I predictor by simply substituting the predictor module while keeping the LIME and SHAP modules unchanged. Using the MHCXAI framework, we successfully generated LIME and SHAP explanation vectors for input peptides from the MHC-Bench dataset across all examined predictors.

In this study, our main emphasis is on providing explanations for input peptides rather than allele sequences, given that not all predictors can process allele input in the form of an amino acid sequence. For instance, NetMHCpan and MHCflurry accept allele inputs as HLA-A02:01 and HLA-A0201, respectively. However, it is important to highlight that our framework has the capability to generate explanations for allele sequences in addition to peptides. This feature can be particularly beneficial for predictors like TransPHLA, which accepts the allele as an input sequence. An illustration of an allele explanation is presented in Supplementary Fig.S2 of Supplementary Note3.

Figure3a illustrates LIME and SHAP explanations generated for all examined predictors for a specific peptide-MHC allele pair (LLVEVLREIHLA-A*02:01). To visualize the explanations, heatmaps are constructed using the attribution values, with lighter colors indicating positive contributions and darker colors indicating little or negative contributions to the binding class. The peptide LLVEVLREI is a binding peptide for the HLA-A*02:01 allele, correctly classified by all MHC class I predictors. However, it is worth noting that the explanations from SHAP and LIME exhibit slight differences. For example, in Fig.3a, both LIME and SHAP attribute high importance to peptide position P2, but SHAP also recognizes peptide position P9 as significant. P9 is typically considered important for binding based on existing literature52,53.

a Examples of SHAP and LIME explanations for all investigated MHC class I predictors for the LLVEVLREI--HLA-A*02:01 pair. To visualize the explanations, the attribution values of positions are used to create heatmaps. For each explanation, a lighter color indicates a positive contribution, while a darker color indicates a smaller or negative contribution to the positive class. SHAP explanations for all the predictors highlight peptide positions P2 and P9 as the most important for binding, while LIME explanations highlight only peptide position P2 as the most important. b LIME and SHAP explanations for NetMHCpan-EL and -BA for the peptide KVAQKQFQL binding to HLA-A*02:01. NetMHCpan-EL classifies KVAQKQFQL correctly, but NetMHCpan-BA does not. SHAP captures these differences in performance and produces different explanations for the two predictor modes. c LIME and SHAP explanations for MHCflurry-PS and -BA for the peptide RVMAPRALL binding to HLA-C*06:02. MHCflurry-PS classifies RVMAPRALL correctly, but MHCflurry-BA does not. Similar to the example in b, SHAP captures these differences in performance and produces different explanations for the two predictor modes. In both NetMHCpan and MHCflurry examples, LIME explanations are unable to indicate positions leading to the difference in prediction outcome.

In Fig.3b, the peptide KVAQKQFQL binds to the MHC allele HLA-A*02:01. However, within NetMHCpan, NetMHCpan-EL correctly predicts it as a binder, while NetMHCpan-BA classifies it as a non-binder. SHAP and LIME explanations were generated for both modes, revealing that SHAP can identify the features (or peptide positions) responsible for the predictions made by NetMHCpan-EL and -BA. For instance, at peptide position P1, the amino acid K positively contributes to the prediction outcome in NetMHCpan-EL, whereas it negatively contributes in NetMHCpan-BA. LIME, on the other hand, produces similar attribution values for both predictors and is unable to highlight the cause for the difference in prediction between the two NetMHCpan modes.

In Fig.3c, the peptide RVMAPRALL is a binder to the MHC allele HLA-C*06:02, classified as a binder by MHCflurry-PS but not by MHCflurry-BA. SHAP and LIME were employed to explore the difference in predictions. SHAP identified that for MHCflurry-PS, peptide positions P1 and P9 play an important role. For MHCflurry-BA, while peptide position P9 is crucial, P1 is not deemed important (refer to the heatmaps in Fig.3c). This distinction between the two explanations helps identify and understand the reasons for different predictions.

These examples demonstrate that XAI techniques can generate explanations for MHC class I predictors. However, explanations produced by distinct XAI techniques for the same predictor may not align. Consequently, we evaluate the validity and quality of LIME and SHAP explanations against XAI metrics.

The average time required to generate an explanation for a single instance is influenced more by the choice of the predictor than the XAI technique itself. For example, generating an explanation (either LIME or SHAP) for MHCfovea takes twice as long as generating a corresponding explanation for MHCflurry.

As stated earlier, global explanations for MHC class I predictors manifest as binding motifs. Determining the binding preference for an MHC allele involves examining the most frequently occurring amino acids at anchor positions, specifically positions 2 and 9 in a 9mer peptide, which are the primary sites responsible for binding to an MHC molecule52,53. This can be extended to other peptide positions, forming a binding motif for an MHC allele.

Biological binding motifs for MHC alleles are generated using experimentally validated strong binders54. With recent MHC class I predictors, binding motifs are derived from peptides predicted as strong binders for a particular allele. These peptides serve as the basis for generating position-specific scoring matrices (PSSM), which are then visually represented as binding motifs. This approach is applied to generate binding motifs for MHCflurry, NetMHCpan, and MHCfovea. By comparing these predictor-generated motifs against the biological motif, it becomes possible to assess whether the predictor has effectively learned the correct binding patterns for a given allele. In this study, we utilize binding motifs from the MHC Motif Atlas database55.

Global explanations may overlook deviations observed in specific inputs. Consider a binding peptide that diverges from the typical biological binding motif pattern. In cases where a predictor correctly classifies this peptide as a binder, it becomes valuable to examine the specific features used by the predictor for this classification. Understanding the input features employed by the predictor for a particular peptide requires a local explanation rather than a binding motif. Specialized patterns like these are difficult to infer with a binding motif.

We illustrate the necessity for local explanations to capture specialized patterns that deviate from biological binding motifs with specific examples. In Fig.4a, the motif for HLA-A*02:01 suggests a preference for amino acids L, I, and, M at anchor position P2, binding to the super hydrophobic B pocket of the MHC molecule. However, despite the unfavorable nature of water-soluble Glutamine (Q) for such a pocket, solved peptide-bound HLA molecule structures indicate that many peptides with Q do bind strongly to HLA-A*02:0156. An example is the peptide FQFICNLLL (see Fig.4a), correctly classified as a binder by the MHCflurry-PS predictor. We generated a local explanation for this peptide using SHAP (see Fig.4a). The highest attribution values were assigned to peptide positions P1, P2, and P9. While the high importance of peptide positions P2 and P9 aligns with their roles as anchor positions, the elevated attribution value for peptide position P1 is rationalized by its crucial role in stabilizing the bound structure, as observed in refs. 53,56,57. It is worth noting that the amino acid Q in position P2 does not appear in the biological binding motif (global explanation) prominently which fails to capture the specialized pattern in this instance.

For a and b, biological binding motifs for HLA-A*02:01 and HLA-A*24:03 are obtained from the MHC Motif Atlas55. For each MHC allele, there are four heatmaps, which are SHAP explanations generated for true positive, true negative, false positive, and false negative peptides predicted using MHCflurry-PS. The peptides in both a and b defy the reasoning for binding based on biological motifs. However, the SHAP explanations are able to highlight the cause behind unexpected outcomes. A lighter color in the explanation heatmap indicates a positive contribution, while a darker color indicates a smaller or negative contribution to the positive class.

In Fig.4a, the true negative instance HQKKNEISF also contains the amino acid Q in position P2, similar to the true positive instance discussed. However, in this case the local explanation shows low attribution values for all other peptide positions. This indicates lack of strong binding signal from those positions, explaining the negative classification.

Figure4b is another example of explanations generated for true positive, true negative, false positive and false negative predictions made by MHCflurry-PS for HLA-A*24:03. In this example, peptide conforming to binding motif is correctly classified as non-binder whereas peptide not conforming to binding motif is correctly classified as binder.

In summary, instance-based explanations are particularly useful in explaining scenarios where binder peptides do not conform to motifs, misclassifications, and understanding a peptide specific pattern used for prediction. Additionally, we show that global explanations can be created using instance-based SHAP and LIME explanations in Supplementary Fig.S3 (in Supplementary Note4) which can be useful for quick comparison across predictors.

To achieve trust in a predictor, the generated explanations for the predictions must be reliable. Assessing the reliability of explanations involves comparing them to ground truth about the input-output relationship, as suggested by prior work such as58,59,60. The ground truth in our case are the residues in the peptide that genuinely contribute to the binding. The contribution of residues (hotspots) can be estimated experimentally using Alanine-scanning mutagenesis61, a resource intensive45 technique. Computationally, this can be achieved using BAlaS45,46 which calculates the difference between the free-energy of binding of original bound complex and mutated bound complex where just one residue of ligand peptide is replaced with alanine. This difference in free-energy of binding is indicated as G and G4.184kJ/mol is considered hot or important residue for binding46. G4.184kJ/mol indicates alanine enhances binding relative to the original residue46. Any value between denotes neutral substitution46. As it is difficult to obtain ground truth for all peptides, we use this G as an independent way of highlighting important residues in the peptide.

First, we compile all the available PDB structures featuring bound peptide-MHC allele complexes as documented in the MHC Motif Atlas55. Subsequently, we refine the list to encompass bound peptides with a length of 9 and narrow down the selection to structures that were consistently classified as binders by all examined MHC class I predictors. The resulting list comprises 250 PDB structures, encompassing 40 distinct MHC alleles (as listed in Supplementary Data5).

We compared the G for these 250 peptide-MHC pairs with the LIME and SHAP explanations generated for all the predictors. The LIME and SHAP values can be positive or negative, similar to G, which indicates residue contribution to the prediction. For each of these 250 peptide-MHC pair we calculated Pearson correlation coefficients between LIME/SHAP explanations and G, for each of the investigated predictors.

In Fig.5a, consider the instance of the peptide ITDQVPFSV bound to HLA-A*02:01, which is correctly classified. BAlaS identifies peptide positions P1, P2, P7, and P9 as hot residues (with G4.184kJ/mol), which are highlighted in red within the peptide-MHC complex. The SHAP explanations, feature red arrows pointing to the positions identified as important by BAlaS. Generally, the models consistently prioritize these positions when making predictions. However, despite having a high G, peptide position P7 is not deemed important by most of the predictors. This suggests that the information from the other three residues is sufficient for the predictors to infer the classification outcome. The distribution of correlation coefficients between SHAP-G and LIME-G (depicted in Fig.5b) indicates a positive correlation between the explanations and the important positions identified by BAlaS. Overall, it is observed that SHAP explanations exhibit a closer correlation compared to LIME explanations.

It is done by comparing the attribution values to the difference in free-energy of binding between wild-type protein-protein interaction and mutated protein-protein interaction, known as G. a For the ITDQVPFSV--HLA-A*02:01 complex, BAlaS highlights that peptide positions P1, P2, P7, and P9 are crucial for binding. Replacing the residues at these positions with alanine leads to an increase in G, indicating instability. This peptide is correctly classified by all the investigated MHC class I predictors, and SHAP explanations are generated for each of them. The explanations mostly match the ground truth, as P1, P2, P7, and P9 (indicated by red arrows) are rightly highlighted as the factors influencing the prediction. b G was calculated for each peptide position in 250 PDB structures containing peptide-MHC allele bound complexes that were correctly classified by all the investigated predictors. The SHAP and LIME explanations correlated positively for most complexes, indicating that the explanations mostly align with the ground truth and can be trusted. The correlation coefficient values are reported in Supplementary Data4. A lighter color in the explanation heatmap indicates a positive contribution, while a darker color indicates a smaller or negative contribution to the positive class.

The observed variance in the distribution of correlation coefficients is not surprising, given that BAlaS G serves as only an approximation of the actual positions involved in binding, and the approach is subject to certain limitations. Notably, the accuracy of the G calculation is influenced by the resolution of the PDB structure (refer to Supplementary Fig.S4 and Supplementary Note5). To address this, we selectively choose PDB structures with the highest resolution when multiple structures are available. Additionally, since G is computed by substituting a residue with alanine, it is challenging to ascertain the contribution of alanine, if present (refer to Supplementary Fig.S5).

Consistency refers to similarity in the explanations produced for two similarly performing predictors on a given input. We assess consistency of an XAI technique by comparing explanations for a given peptide between two similarly performing MHC class I predictors (Fig.1c).

To select two predictors with comparable performance, we choose the top two predictors from our results in Section 2 (see Fig.2a), namely MHCflurry-PS and MHCfovea. Additionally, the AUROC scores for these two predictors exhibit a high correlation, as indicated in Fig.6c, demonstrating substantial similarity in their performance.

a For an input peptide, explanations were generated for MHCflurry-PS and MHCfovea using SHAP and LIME. Pearson correlation coefficients were calculated between these explanations, and the process was repeated for 200 input peptides for each of the alleles presented in the plot. The distribution of Pearson correlation coefficients is closer to one, indicating high similarity between the two explanations for the two predictors on the same input. The correlation coefficient values are reported in Supplementary Data6. b In addition to the Pearson correlation coefficient, Euclidean distances were calculated between two explanations for two predictors on the same input. For Euclidean distance, values closer to zero indicate high similarity and high consistency. c Correlation heatmap for AUROC scores between investigated MHC class I predictors. MHCflurry-PS and MHCfovea are highly correlated in their performances.

We selected 9 alleles (3 each from HLA-A, B and C) and for each allele, we randomly selected 200 peptides from our MHC-Bench dataset to generate local explanations, independently using each of SHAP and LIME, for MHCfovea and MHCflurry-PS. For both LIME and SHAP, to compare the similarity between the explanations from the two predictors, we computed Pearson correlation and Euclidean distance.

In Fig.6a, the distribution of Pearson correlations between explanations generated individually for MHCflurry-PS and MHCfovea using LIME and SHAP is presented for all nine alleles. Overall, the majority of the SHAP and LIME explanations exhibited high correlation. In Fig.6b, the distribution of Euclidean distances between the explanations of the two predictors is presented. Explanations that are similar will have a Euclidean distance closer to zero. It is noteworthy that the Euclidean distance distribution for LIME has a narrow range and tends to be closer to zero compared to SHAP. This observation suggests that LIME produces more consistent explanations compared to SHAP.

We also created a baseline distance between explanations for the two predictors using the following procedure. First, we generated 100 random explanations for each original MHCfovea explanation by randomly permuting the attribution values. Next, we calculated the distance between each of these 100 random explanations and the original MHCflurry-PS explanation. The baseline distance was then computed by averaging these 100 distances. This process was repeated for all 200 peptides chosen per allele. Consequently, we obtained 200 Euclidean distances between the original MHCflurry-PS and MHCfovea explanations, along with their corresponding baseline distances. We compared these two distributions for each allele. We confirmed that the two distributions for both SHAP and LIME were statistically different using Kruskal-Wallis test at 5% significance level. The p-value, H-statistics and effect size are reported in Supplementary Tables1 and 2 for SHAP and LIME respectively. It is also worth noting that the Euclidean distance for both LIME and SHAP explanations were smaller than the corresponding average baseline Euclidean distance for nearly all the input peptides (99% input peptides).

We confirmed that the two distributions for both SHAP and LIME were statistically different using Kruskal-Wallis test at 5% significance level. The p-value, H-statistics, and effect size are reported in Supplementary Tables1 and 2 for SHAP and LIME, respectively. Additionally, it is noteworthy that the Euclidean distances for both LIME and SHAP explanations were smaller than the corresponding average baseline Euclidean distances for nearly all input peptides (99% of input peptides).

Stability of an explanation technique refers to the extent to which explanations for similar inputs (with same output) over a given predictor are close. We use the MHCflurry-PS predictor to asses stability of the LIME and SHAP techniques, independently. To identify input peptides that are similar, we perform clustering over a subset of peptide sequences for HLA-A*02:01. Using GibbsCluster-2.062,63, we cluster the peptides into 110 clusters. The number of clusters that yields highest average KullbackLeibler Distance (KLD) is considered to be the optimum number of clusters. We found choosing 10 clusters has the highest KLD with cluster size ranging between 7001000 peptides. The plot showing KLD distribution and cluster motifs generated from GibbCluster is provided in Supplementary Fig.S6. Peptides within a cluster are considered similar.

From each of these clusters, we sampled 100 peptides that are binders and generated explanations for these peptides. We calculated the Euclidean distance between all pairs of peptides within each cluster, and this is referred to as the intracluster distance distribution. As a comparison, we also computed the distance between explanations for peptides from different clusters, referred to as Intercluster distance. We show results for the top six most unrelated cluster pairs (c2, c5), (c3, c5), (c3, c8), (c5, c6), (c5, c9), (c5, c10), based on the similarity of their position-specific scoring matrix, in Fig.7.

a Euclidean distance distribution for the top six cluster pairs using LIME. For each pair, there are three distributions - IntraclusterL, Intercluster, and IntraclusterR. For any two pairs (e.g., c2, c5), the intracluster explanation distance distribution for the left cluster (c2) and right cluster (c5) are IntraclusterL and IntraclusterR, while Intercluster is the distribution of explanation distances between the two clusters. b Euclidean distance distribution for the top six cluster pairs using SHAP. For each pair, there are three distributions - IntraclusterL, Intercluster, and IntraclusterR. For any two pairs (e.g., c2, c5), the intracluster explanation distance distribution for the left cluster (c2) and right cluster (c5) are IntraclusterL and IntraclusterR, while Intercluster is the distribution of explanation distances between the two clusters.

For each cluster pair in Fig.7, we have three distributions: IntraclusterL, Intercluster, and IntraclusterR. Consider the pair (c2, c5), where IntraclusterL represents the intracluster Euclidean distance distribution for cluster c2 (or the left cluster), Intercluster is the intercluster Euclidean distance distribution between c2 and c5, and IntraclusterR is the intracluster Euclidean distance distribution for cluster c5 (or the right cluster). The notation of left-right for Intracluster is arbitrary. It is worth noting that intracluster distances are lower than the intercluster distances, indicating that LIME and SHAP explanations for peptides within the same cluster are more similar, suggesting stability of explanations. We confirmed that the differences between intracluster and intercluster distance distributions in Fig.7 are statistically significant (KruskalWallis test). The p-value, H-statistics, and effect size are reported in Supplementary Table3.

Read more:
Building trust in deep learning-based immune response predictors with interpretable explanations | Communications ... - Nature.com